帮助我在Python中使用我的backprop实现

时间:2010-10-21 13:59:17

标签: python algorithm math neural-network

EDIT2:

新的训练集......

输入:

[
 [0.0, 0.0], 
 [0.0, 1.0], 
 [0.0, 2.0], 
 [0.0, 3.0], 
 [0.0, 4.0], 
 [1.0, 0.0], 
 [1.0, 1.0], 
 [1.0, 2.0], 
 [1.0, 3.0], 
 [1.0, 4.0], 
 [2.0, 0.0], 
 [2.0, 1.0], 
 [2.0, 2.0], 
 [2.0, 3.0], 
 [2.0, 4.0], 
 [3.0, 0.0], 
 [3.0, 1.0], 
 [3.0, 2.0], 
 [3.0, 3.0], 
 [3.0, 4.0],
 [4.0, 0.0], 
 [4.0, 1.0], 
 [4.0, 2.0], 
 [4.0, 3.0], 
 [4.0, 4.0]
]

输出:

[
 [0.0], 
 [0.0], 
 [0.0], 
 [0.0], 
 [0.0], 
 [0.0], 
 [0.0], 
 [0.0], 
 [0.0], 
 [0.0], 
 [0.0], 
 [0.0], 
 [0.0], 
 [0.0], 
 [0.0], 
 [0.0], 
 [0.0], 
 [0.0], 
 [1.0], 
 [1.0], 
 [0.0], 
 [0.0], 
 [0.0], 
 [1.0], 
 [1.0]
]

EDIT1:

我已使用我的最新代码更新了问题。我解决了一些小问题,但在网络学习之后,我仍然为所有输入组合获得相同的输出。

以下是backprop算法:Backprop algorithm


是的,这是一个功课,要在开始时明确这一点。

我应该在一个简单的神经网络上实现一个简单的反向传播算法。

我选择Python作为此任务的首选语言,我选择了这样的神经网络:

3层:1个输入,1个隐藏,1个输出层:

O         O

                    O

O         O

两个inptut神经元都有一个整数,输出神经元上有1或0。

这是我的整个实现(有点长)。 Bellow it我将选择更短的相关片段,我认为错误可能位于:

import os
import math
import Image
import random
from random import sample

#------------------------------ class definitions

class Weight:
    def __init__(self, fromNeuron, toNeuron):
        self.value = random.uniform(-0.5, 0.5)
        self.fromNeuron = fromNeuron
        self.toNeuron = toNeuron
        fromNeuron.outputWeights.append(self)
        toNeuron.inputWeights.append(self)
        self.delta = 0.0 # delta value, this will accumulate and after each training cycle used to adjust the weight value

    def calculateDelta(self, network):
        self.delta += self.fromNeuron.value * self.toNeuron.error

class Neuron:
    def __init__(self):
        self.value = 0.0        # the output
        self.idealValue = 0.0   # the ideal output
        self.error = 0.0        # error between output and ideal output
        self.inputWeights = []
        self.outputWeights = []

    def activate(self, network):
        x = 0.0;
        for weight in self.inputWeights:
            x += weight.value * weight.fromNeuron.value
        # sigmoid function
        if x < -320:
            self.value = 0
        elif x > 320:
            self.value = 1
        else:
            self.value = 1 / (1 + math.exp(-x))

class Layer:
    def __init__(self, neurons):
        self.neurons = neurons

    def activate(self, network):
        for neuron in self.neurons:
            neuron.activate(network)

class Network:
    def __init__(self, layers, learningRate):
        self.layers = layers
        self.learningRate = learningRate # the rate at which the network learns
        self.weights = []
        for hiddenNeuron in self.layers[1].neurons:
            for inputNeuron in self.layers[0].neurons:
                self.weights.append(Weight(inputNeuron, hiddenNeuron))
            for outputNeuron in self.layers[2].neurons:
                self.weights.append(Weight(hiddenNeuron, outputNeuron))

    def setInputs(self, inputs):
        self.layers[0].neurons[0].value = float(inputs[0])
        self.layers[0].neurons[1].value = float(inputs[1])

    def setExpectedOutputs(self, expectedOutputs):
        self.layers[2].neurons[0].idealValue = expectedOutputs[0]

    def calculateOutputs(self, expectedOutputs):
        self.setExpectedOutputs(expectedOutputs)
        self.layers[1].activate(self) # activation function for hidden layer
        self.layers[2].activate(self) # activation function for output layer        

    def calculateOutputErrors(self):
        for neuron in self.layers[2].neurons:
            neuron.error = (neuron.idealValue - neuron.value) * neuron.value * (1 - neuron.value)

    def calculateHiddenErrors(self):
        for neuron in self.layers[1].neurons:
            error = 0.0
            for weight in neuron.outputWeights:
                error += weight.toNeuron.error * weight.value
            neuron.error = error * neuron.value * (1 - neuron.value)

    def calculateDeltas(self):
        for weight in self.weights:
            weight.calculateDelta(self)

    def train(self, inputs, expectedOutputs):
        self.setInputs(inputs)
        self.calculateOutputs(expectedOutputs)
        self.calculateOutputErrors()
        self.calculateHiddenErrors()
        self.calculateDeltas()

    def learn(self):
        for weight in self.weights:
            weight.value += self.learningRate * weight.delta

    def calculateSingleOutput(self, inputs):
        self.setInputs(inputs)
        self.layers[1].activate(self)
        self.layers[2].activate(self)
        #return round(self.layers[2].neurons[0].value, 0)
        return self.layers[2].neurons[0].value


#------------------------------ initialize objects etc


inputLayer = Layer([Neuron() for n in range(2)])
hiddenLayer = Layer([Neuron() for n in range(100)])
outputLayer = Layer([Neuron() for n in range(1)])

learningRate = 0.5

network = Network([inputLayer, hiddenLayer, outputLayer], learningRate)

# just for debugging, the real training set is much larger
trainingInputs = [
    [0.0, 0.0],
    [1.0, 0.0],
    [2.0, 0.0],
    [0.0, 1.0],
    [1.0, 1.0],
    [2.0, 1.0],
    [0.0, 2.0],
    [1.0, 2.0],
    [2.0, 2.0]
]
trainingOutputs = [
    [0.0],
    [1.0],
    [1.0],
    [0.0],
    [1.0],
    [0.0],
    [0.0],
    [0.0],
    [1.0]
]

#------------------------------ let's train

for i in range(500):
    for j in range(len(trainingOutputs)):
        network.train(trainingInputs[j], trainingOutputs[j])
        network.learn()

#------------------------------ let's check


for pattern in trainingInputs:
    print network.calculateSingleOutput(pattern)

现在,问题在于,在学习网络后,对于所有输入组合,似乎返回的浮点数非常接近0.0,即使是那些应该接近1.0的浮点数。

我在100个周期内训练网络,在每个周期中我都会这样做:

对于训练集中的每组输入:

  • 设置网络输入
  • 使用sigmoid函数计算输出
  • 计算输出图层中的错误
  • 计算隐藏图层中的错误
  • 计算权重'增量

然后我根据学习率和累积的增量来调整权重。

这是我对神经元的激活功能:

def activationFunction(self, network):
    """
    Calculate an activation function of a neuron which is a sum of all input weights * neurons where those weights start
    """
    x = 0.0;
    for weight in self.inputWeights:
        x += weight.value * weight.getFromNeuron(network).value
    # sigmoid function
    self.value = 1 / (1 + math.exp(-x))

这是我如何计算增量:

def calculateDelta(self, network):
    self.delta += self.getFromNeuron(network).value * self.getToNeuron(network).error

这是我的算法的一般流程:

for i in range(numberOfIterations):
    for k,expectedOutput in trainingSet.iteritems():
        coordinates = k.split(",")
        network.setInputs((float(coordinates[0]), float(coordinates[1])))
        network.calculateOutputs([float(expectedOutput)])
        network.calculateOutputErrors()
        network.calculateHiddenErrors()
        network.calculateDeltas()
    oldWeights = network.weights
    network.adjustWeights()
    network.resetDeltas()
    print "Iteration ", i
    j = 0
    for weight in network.weights:
        print "Weight W", weight.i, weight.j, ": ", oldWeights[j].value, " ............ Adjusted value : ", weight.value
        j += j

输出的最后两行是:

0.552785449458 # this should be close to 1
0.552785449458 # this should be close to 0

它实际上返回所有输入组合的输出数。

我错过了什么吗?

1 个答案:

答案 0 :(得分:5)

看起来你得到的几乎是神经元的初始状态(差不多self.idealValue)。也许你不应该在提供实际数据之前初始化这个神经元?

编辑:好的,我在代码中看起来更深一些并简化了一下(将在下面发布简化版)。基本上你的代码有两个小错误(看起来像你忽略的东西),但这会导致网络无法正常工作。

  • 您在学习阶段忘记在输出图层中设置expectedOutput的值。没有它,网络肯定无法学习任何东西,并将始终坚持初始idealValue。 (这是我在一读时发现的行为)。甚至可以在您对训练步骤的描述中发现这一点(如果您没有发布代码,可能会发现这一点,这是我知道实际发布代码隐藏错误而非制作错误的罕见情况之一明显)。你在EDIT1之后修好了这个。
  • 在calculateSingleOutputs中激活网络时,您忘记激活隐藏图层。

显然,这两个问题中的任何一个都会导致一个不连贯的网络。

一旦纠正,它就可以了(好吧,它在我的代码的简化版本中)。

错误并不容易发现,因为初始代码太复杂了。在引入新类或新方法之前,您应该三思而后行。没有创建足够的方法或类会使代码难以阅读和维护,但创建太多可能会使其更难以阅读和维护。你必须找到合适的平衡点。我找到这种平衡的个人方法是遵循code smells并在他们引导我的地方重构技巧。有时添加方法或创建类,有时会删除它们。它当然不是完美的,但这对我有用。

在应用了一些重构之后,下面是我的代码版本。我花了大约一个小时更改你的代码,但始终保持其功能相当。我认为这是一个很好的重构练习,因为最初的代码真的很糟糕。在重构之后,它花了5分钟来发现问题。

import os
import math

"""
A simple backprop neural network. It has 3 layers:
    Input layer: 2 neurons
    Hidden layer: 2 neurons
    Output layer: 1 neuron
"""

class Weight:
    """
    Class representing a weight between two neurons
    """
    def __init__(self, value, from_neuron, to_neuron):
        self.value = value
        self.from_neuron = from_neuron
        from_neuron.outputWeights.append(self)
        self.to_neuron = to_neuron
        to_neuron.inputWeights.append(self)

        # delta value, this will accumulate and after each training cycle
        # will be used to adjust the weight value
        self.delta = 0.0

class Neuron:
    """
    Class representing a neuron.
    """
    def __init__(self):
        self.value = 0.0        # the output
        self.idealValue = 0.0   # the ideal output
        self.error = 0.0        # error between output and ideal output
        self.inputWeights = []    # weights that end in the neuron
        self.outputWeights = []  # weights that starts in the neuron

    def activate(self):
        """
        Calculate an activation function of a neuron which is 
        a sum of all input weights * neurons where those weights start
        """
        x = 0.0;
        for weight in self.inputWeights:
            x += weight.value * weight.from_neuron.value
        # sigmoid function
        self.value = 1 / (1 + math.exp(-x))

class Network:
    """
    Class representing a whole neural network. Contains layers.
    """
    def __init__(self, layers, learningRate, weights):
        self.layers = layers
        self.learningRate = learningRate    # the rate at which the network learns
        self.weights = weights

    def training(self, entries, expectedOutput):
        for i in range(len(entries)):
            self.layers[0][i].value = entries[i]
        for i in range(len(expectedOutput)):
            self.layers[2][i].idealValue = expectedOutput[i]
        for layer in self.layers[1:]:
            for n in layer:
                n.activate()
        for n in self.layers[2]:
            error = (n.idealValue - n.value) * n.value * (1 - n.value)
            n.error = error
        for n in self.layers[1]:
            error = 0.0
            for w in n.outputWeights:
                error += w.to_neuron.error * w.value
            n.error = error
        for w in self.weights:
            w.delta += w.from_neuron.value * w.to_neuron.error

    def updateWeights(self):
        for w in self.weights:
            w.value += self.learningRate * w.delta

    def calculateSingleOutput(self, entries):
        """
        Calculate a single output for input values.
        This will be used to debug the already learned network after training.
        """
        for i in range(len(entries)):
            self.layers[0][i].value = entries[i]
        # activation function for output layer
        for layer in self.layers[1:]:
            for n in layer:
                n.activate()
        print self.layers[2][0].value


#------------------------------ initialize objects etc

neurons = [Neuron() for n in range(5)]

w1 = Weight(-0.79, neurons[0], neurons[2])
w2 = Weight( 0.51, neurons[0], neurons[3])
w3 = Weight( 0.27, neurons[1], neurons[2])
w4 = Weight(-0.48, neurons[1], neurons[3])
w5 = Weight(-0.33, neurons[2], neurons[4])
w6 = Weight( 0.09, neurons[3], neurons[4])

weights = [w1, w2, w3, w4, w5, w6]
inputLayer  = [neurons[0], neurons[1]]
hiddenLayer = [neurons[2], neurons[3]]
outputLayer = [neurons[4]]
learningRate = 0.3
network = Network([inputLayer, hiddenLayer, outputLayer], learningRate, weights)

# just for debugging, the real training set is much larger
trainingSet = [([0.0,0.0],[0.0]),
               ([1.0,0.0],[1.0]),
               ([2.0,0.0],[1.0]),
               ([0.0,1.0],[0.0]),
               ([1.0,1.0],[1.0]),
               ([2.0,1.0],[0.0]),
               ([0.0,2.0],[0.0]),
               ([1.0,2.0],[0.0]),
               ([2.0,2.0],[1.0])]

#------------------------------ let's train
for i in range(100): # training iterations
    for entries, expectedOutput in trainingSet:
        network.training(entries, expectedOutput)
    network.updateWeights()

#network has learned, let's check
network.calculateSingleOutput((1, 0)) # this should be close to 1
network.calculateSingleOutput((0, 0)) # this should be close to 0

顺便说一下,还有第三个问题我没有纠正(但很容易纠正)。如果x太大或太小(> 320或<-320)math.exp()将引发异常。如果你申请训练迭代,比如几千,就会发生这种情况。我所看到的最简单的纠正方法是检查x的值,如果它太大或太小,将Neuron的值设置为0或1,具体取决于具体情况,即极限值。